About Iris

A creative director that sees, listens, and creates

Iris is a production-grade multimodal AI studio built for AI Engineer portfolios. It combines voice input, live vision, LLM tool calling, image generation, and TTS narration in a single streaming pipeline — with real cost and latency observability.

Demo mode runs entirely on Vercel without API keys: scripted orchestration, browser TTS, and placeholder imagery. Add keys locally for live GPT-4o, Whisper, DALL-E, and SDXL.

Architecture

Server-Sent Events stream each pipeline stage to the client in real time.

flowchart LR
  User[Voice + Vision] --> Studio[Next.js Studio]
  Studio --> ASR[Whisper ASR]
  Studio --> SSE[SSE Stream API]
  SSE --> Orch[LLM Orchestrator]
  Orch --> Tools[Tool Calling]
  Tools --> Img[SDXL / DALL-E]
  Orch --> TTS[TTS Narration]
  Studio --> Obs[Observability]

Mermaid diagram (render on GitHub README). Tools: generate_image, edit_image, describe_scene.

Streaming

SSE events for thinking, tool calls, images, and responses

Demo mode

Full UX without API keys — ideal for portfolio reviewers

Observability

Per-session cost, P50/P95 latency, stage breakdown

Replays

Shareable session snapshots with timeline playback

Built with Next.js, TypeScript, OpenAI, and Replicate. Deployed on Vercel with GitHub Actions CI.

Try the Studio View Observability Engineering Deep-Dive